Source Camera Classification and Clustering from Sensor Pattern Noise - Applied to Child Sexual Abuse Investigations

Abstract

In police investigations concerning child sexual abuse crimes, the most important evidence is often digital images. Due to the large number of images in such cases, tools are needed to reduce the amount of manual work. A desirable feature currently not available to law enforcement personnel is the ability to reliably tell if an image was captured by a specific camera or not. Another useful feature is the ability to cluster images based on the source camera. These goals can be achieved by the use of Sensor Pattern Noise and in this study methods to extract the noise from images will be evaluated. Furthermore different clustering methods and new clustering heuristics, such as pre-clustering based on camera model, is evaluated. To improve clustering results correlation between reference patterns constructed from already clustered images is studied. The evaluation of the denoising algorithms concluded that the color decoupled version of the Mihçak denoising filter was superior to the other tested methods. The correlation between reference patterns from clusters of images was concluded to be highly dependent on the number of images in the clusters. The introduction of pre-clustering based on if two images where from the same camera, using features from the image and noise and a trained classifier, decreased the time consumption of the clustering algorithms considerably, thus making the clustering methods more feasible when the amount of images is large. By merging the noise from clusters into reference patterns more images were grouped together than when only single image noise patterns were compared to each other.